Privacy, safety, and AI mistakes
The canonical version always lives at https://tavina.ai/docs/friendly/privacy-safety/privacy-overview
Privacy, safety, and AI mistakes
Tavina is designed to be useful, but it is still an AI assistant. AI can make mistakes, omit details, misunderstand your request, or give an answer that sounds more certain than it should.
Use Tavina responsibly and review important work before relying on it.
AI output is not guaranteed
Tavina may produce incorrect, incomplete, outdated, or misleading information. You remain responsible for deciding whether an answer is safe to use.
Use extra care with:
- medical or health decisions;
- legal matters;
- financial or tax decisions;
- safety-critical instructions;
- employment or business-critical decisions;
- account access, payments, or sensitive records.
For those topics, treat Tavina as a drafting or brainstorming assistant, not as a professional advisor.
Tavina also watches for prompt-injection attempts โ text inside files, pages, or messages that tries to steer your agent into ignoring your instructions. When something looks suspicious, your agent sees a warning explaining what was flagged and can tell you about it; the warning does not block the conversation, and you can judge for yourself what to do with the flagged content. These notices are on for every account (tavina/tavina#2157).
Protect sensitive information
Avoid sharing secrets or unnecessary sensitive information with Tavina. Do not enter passwords, payment card numbers, private keys, one-time codes, or credentials.
When uploading documents, make sure you have permission to use them with Tavina. When connecting apps, choose the lowest trust level that works for your task.
Files and Knowledge Library privacy expectations
Document-like uploads may become Knowledge Library Sources with extracted text, chunks, metadata, and retention state. Original files are not necessarily retained forever, and retention should depend on safe extraction and explicit controls.
Photo uploads can contain image metadata such as capture date, camera/device, lens settings, dimensions, orientation, and GPS/location coordinates when present. Tavina may use a bounded summary of available image metadata to improve visual analysis. For Tavina accounts, available image metadata โ including GPS/location โ is shared with analysis by default because location can be useful for real-world image questions. If you do not want location metadata included, use any available account or environment opt-out/redaction controls before upload. Metadata may also be absent, stripped by another app, inaccurate, or intentionally disabled.
If a Source is failed, partial, OCR-needed, encrypted/blocked, or missing artifacts, Tavina should say so instead of pretending it has complete access.
Memory expectations
Memory may help Tavina personalize future work, but memory is not a perfect official record. Correct Tavina when it recalls something wrong, and do not use memory for secrets or high-stakes records.
Consent and action safety
If Tavina asks whether you want it to do something, silence is not approval. A follow-up reminder or background continuation should re-surface the question or wait for your answer instead of treating Tavina’s own suggestion as permission.
Tavina may be more or less proactive depending on your Identity or Autonomy setting, but that setting is not a blanket authorization. Higher autonomy can help Tavina continue clear, low-risk commitments; it should not bypass confirmation for destructive, irreversible, payment-related, externally visible, or high-blast-radius actions.
Connected-app safety
Connected apps may expose or change information outside Tavina. Review trust levels carefully and check the connected app itself when an action matters. Trust and autonomy settings do not remove the need to confirm sensitive or high-impact actions.
What this page does not replace
This page is customer guidance. It does not replace formal privacy policy wording, legal terms, billing terms, or support commitments. Those items need their own approved public sources.